Deep Learning Computer Vision Interview Series #4-Asked In Interview ⭐ ⭐⭐⭐⭐⭐⭐⭐

[@visintel]

Moving the camera is equivalent to applying a perspective transformation. If we can get the transformation, we can reverse it and do the inference on a simulated pose 1 image. 1. Get a background-only image from camera pose 1 and 2 2. Calculate the homography between images by looking at sift feature matches 3. Apply the inverse homography on new views to make them match the first camera pose 4. Use the same model to do inference. If the change in location introduces changes in intensity/hue/saturation, we can also do histogram matching to correct for that.

[@damodharratnamthappeta2022]

since the model retraining is not possible,may be changing the lens which has more focus and then it will be able to capture it from that height

[@peterr2237]

The simplest answer would be if the existing model is orientation invariant, there is nothing to change. Otherwise, a certain orientation compensation needs to be done such as homography, or perspective transformation.

[@ganapatilmillenial]

Using calibration matrices(intrinsic+extrinsic parameter) we can transform image from new(new height) coordinate system to old (old position height) coordinate system..

[@akashchoudhuri3932]

Sir, please start a Computer-Vision Playlist

[@saravanans9895]

By including some sort of lens that would compensate for the reduction in the size of the car due to the change in height?

[@vivekkandeyang6175]

We can try preprocessing the image such that the distribution at the new position matches the old. Though preprocessing will not fix the things completely but we will definitely get closer results.

[@d3v487]

1. Change the threshold 1st. 2. Use deep sort or you can say a tracker, integrate it to your model (pertained version easily available)

[@sivasai8829]

If we kept the camera higher from original position the object size will be decreased and some times our model may not recognize it.Then i think we can zoom the camera like appearimg from old position then it may work 🤔 and of course we should use good resolution camera

[@mouryasashank2213]

This problem is called Concept Drift or Data Drift. It is solved by Periodically Re-Fit, Periodically Update, Weight Data, Learn The Change, Detect and Choose Model, Data Preparation.

[@Diya-vs4ym]

Apply morphology process

[@kaiwalyagondchar8336]

Firstly we calculate the Max distance between camera and the object at which model will perform efficiently using stereo vision and then we can add a constraint that the object detection will work within that perticular distance.

[@gauravchaudhari618]

I got IISC Mtech AI offer can you please take interview with student or review of the course I am comparing with MTech cs in other IIT's

[@sumannelli1427]

We can use object tracking concept...there which pixels are moving we need to scaleup those part..thanks

[@zeeshansheikh8163]

By Considering some of the Post-processing techniques. 1. Based on the height w.r.t aspect ratio calculating the original ratio of area of interest and we can crop it then zoom the image to get the original region. (here while cropping we can exclude the extra region which is in focus and which is not useful). note: why crop then zoom? why not zoom and crop? --> after zooming and then cropping we may loose some pixel values so the best approach is to crop then zoom. 2. Even then we are facing any blur or noise in the image we can use the below techniques to enhance the image. --> Sharpening the image and smoothening the values with different kernel sizes. --> Using Image Enhancement techniques or point processing techniques based on contrast. 3. Or else if we have a lens which is cable of getting the focus region from the lens.(Don't know much about iot integration with respect to lens)

[@tejach8279]

The major change would the change in perspective of the camera We can use homography to obtain the images similar to the previous position of the camera

[@dunnasuryanarayana5918]

Increasing height might make the objects look smaller, hence slice the image in to equal sized parts(May be like a grid) and resize the image according to the input size. Detect the objects in all those sliced images and finally concatenate the predictions.

[@somnathroy475]

One can use SAHI Slicing to predict the outputs, basically it slices the image in overlapping windows with a prefixed stride and calculate model prediction for each of the slices then combines the results applying IOU to get the final results.

[@ArunKumar-tu8ln]

We can take the ratio of image height and width with respect to height of camera, hence the change in height of camera will be compensated as ratio is taken, instead of original.

[@ramagupta5192]

@krish Naik, Can you please tell the answer? or maybe a hint?

[@raviyadav2552]

sir pls make a series on computer vision.

[@pr0t0n86]

If the camera position is slightly changed then there won't be much of an issue at the time of inference because if YOLO V3+ architecture is used then to some extent it internally takes care of images with different perspective like near to the camera or far away from it.

[@shubhamkanwal8977]

We can experiment with adjusting threshold for the model . Previously as camera was closer so it would have been making predictions with higher accuracy so threshold value would have been high > 80%. If we try to reduce threshold to some extent again experimental, we can compensate for increased distance. till the model is re-trained with augmented datasets using rotation, mosaic , filters augmentation

[@prafulbs7216]

You can use Deep sort algorithm.

[@sharifm3682]

So when we change the camera position, and run our model on new images, it will produce less accuracy in object detection on new images per new height, so here simplest and cost effective approach would be data "pre-processing", find the difference between higher accuracy data and current image (from new height w/ less accuracy) and then project the new image by that difference. Its perfect example of data drift/shift. Since training the model is expensive, better pre-process the new data by finding the diff between old (high accuracy) & new (less accuracy) data. Since new data has more noise compare to old data (which was perfect) , so we have to minimize it or completely get rid of it thru pre-processing.

[@MitranRM]

is the answer is perspective transform

[@prasannavenkatesh9700]

After changing the angle it may affect on detection accuracy. We can try the affine transformation of the image before passing to the model. It may help but without seeing the exact scenario it's hard for drawing the solutions. Similar problem I have faced. I have to detect the keypoints of the person but the camera view was very much tilted so before passing the frames to the model I've shifted the frame to -40 degree. After rotating with -40 degree the results were good. Very much eager to see the approach

[@asqwer453]

I'd focus on creating an argument parser and set my threshold to x as the default. (70%). From their, if I have to relocate the camera to a higher point, and then readjust the threshold within my argument parser to a lower number (40-50% confidence) and see how the model now performs. I'd focus on first ensuring the model localizes the object, then focus on the classification of that object until my model gets a close enough output that I desire.

[@vikankshnath8068]

1. as the size of the objects in the view area of camera will decrease with the height/distance of the camera from the scene, so a simple solution is to do scale up the image propotianally and feed up that approximate cropped area which was previously in camera view. 2. Now, the scaling up images may reduce the quality of images so go with such scaling approaches that does not harm image quality, and add some deblurring filters. Moreover, if budget allows replace the camera at new height with better resolution camera so that after scaling up the images, image quality remains same as that of previous ones.

[@nuzhat_tasfia3431]

Hello Krish. I am a beginner in computer vision. It's my humble request to make a computer vision playlist from scratch in google colab. Please Krish its a request. We really need it. No one can explain beautifully as you do. Please Krish..🙏🙏

[@BharathKumarThota-eg8jc]

i feel we can change the coordinate system, however i feel there is no need of retraintng the model as we train the model in large data sets.mostly this cases may cover up.

[@krishnamore2281]

If the camera is placed at a height we will need to change the view by tilting the view.so the model will be of no use as it is trained with different view. But to tackle this problem we can move the camera behind at the elevated height .

[@akash_thing]

May be by specifying the region of interest.

[@HarshSingh-zp7jb]

Well, now that you have placed it a different height, the model will have problems related to Scaling, visibility (since now the objects are smaller), one possible solution is, before we pass the image we can perform Data Augmentation techniques like zooming and scaling, etc.

[@akramahmad7541]

All trainable weights in all connection nodes of the whole modified model could fine-tuned using new camera location data. In this fine-tuning process, a small learning rate (learning rate = 10− 5) could used to make a small variation to the pre-trained parameters. In this way, we will preserve the valuable parameters as much as possible by avoiding dramatic changes on the pre-trained parameters and let the model learn the special characteristics of changed camera position data

[@veritasiumop3900]

He said no retraining

[@debarchanbasu1962]

Could the camera be set to some zoom? That way, the height gets compensated for.

[@sriramswaminathan8633]

The dimensions of the object will reduce proportionally with the height. Features taken from the image captured at the new height can therefore be multiplied by the height as the factor, to detect/classify the objects.

[@afaqahmad1816]

Good way but what about the perspective transformation of an image, as we already know about height and earlier position?

[@muhammadsalmanhassan7544]

Hello sir kirish please make complete playlist on Open CV like Deeplearning hope you are listening thanks love from Pakistan

[@crpatel9090]

I have been asked one question: Company background : The company is working on a product for Ai based face detection security in United states Stadium and large venues. Question : we have a trained face detection algorithm, it is working on most of the faces but the model has a hard time dealing with people with long nose ( this type of face). Collecting data is not feasible because it is a privacy issue and the United States have strict low on how you use and collect face data. How would you solve this or identify it.

[@balajichippada]

We can try to edit or augment our images in the dataset to elongate the noses and train tthe model on these augmented dataset.

[@chauhandharmraj3536]

@@balajichippada exactly right brother ... Now a days aug. Is everything 😂

[@vishaljaiswar9447]

True. Data augmentation would be feasible. One of the way would be augmenting someone else's long nose on another person's face. It would be tricky, but surely possible.

[@crpatel9090]

@@chauhandharmraj3536 May be one of the reason why many big companies AI system fails in fairness ( example Twitter Crop )

[@chauhandharmraj3536]

@@crpatel9090 can you pls explain.. I am not much aware about this topic..

[@Shivakrishnakarnati]

Model will be unable to detect the objects properly, since the distance is increased. In this case, we might face False Negative issue. My solution would be decreasing the confidence threshold for detection. We can also give the original image size, without resizing it so that model can detect well.

[@oleksander]

We should change the angle of the camera. The main issue here is not the distance, but rotation

[@pareesepathak7348]

We can use homography transformation to images obtained and transform images to the previous location where our detector is trained.

[@soumyaranjansethi1790]

Sir it's a request kindly please create separate videos for your answer regarding the problems it would be really helpful

[@AshishKumar-ox2ng]

Roadmap for object detection

[@ishtiakahmed3272]

Hello sir, I have completed your machine learning and deep learing playlist and understood them. Do I need to learn additionally for deep learning or your playlist is enough? Please answer my query. Thanks for your easily explained video.

[@krishnaik06]

U can start with the playlist

[@ishtiakahmed3272]

@@krishnaik06 which playlist?

[@nithinmamidala]

@@ishtiakahmed3272 krish naik live project exection playlist

[@AM-db7jp]

Perspective Distortion, Resolution and Detail Loss Fix: Image preprocessing Adjustments Perspective Transformation

[@debabratapattnayak5442]

Plz make a vedio on AIvsMLvsDEELEARNINGvsData science

[@kishlayamourya3141]

kzbin.info/www/bejne/oWOzkKN-hqmfppI

[@rohitprakashsahoo4004]

Scale the image obtain by camera at program like factor 2times or three time. then use it for further. No need to train once again.

[@GowrishankarV]

We assume a right ANGLE triangle scenario here with new camera inputs giving us the hypotenuse. We can compensate with projection of the new angle on x axis. Multiply every pixel with cos(theta). Theta being the angle change at the point of focus.

[@adityavijaysomani2546]

attaboy

[@rupeshpatil6636]

By changing the filters..

[@digidim]

We can try doing a Birds-eye-view transformation on the resulting image before passing it to the model

[@vibhasanthanam4493]

Employ video processing techniques such as zooming into the frames captured and feed it to the algorithm.

[@thesleepyhead7273]

is that your picture with your own quote in the back , and ffs please keep camera a bit far from your face

[@thirumaleshbr1131]

Will it work if we zoom the lens to requred extent ?

[@sohailquadri5456]

Bro u would hav answer the question too bcoz we have to search it Iam doing a working project on object detection using open cv can you pls help me

[@prashantbhat8616]

Apply some image processing algorithms to compensate for the object resolution lost due to height, like zoom in and then perform super resolution or some other upscaling techniques before feeding the images to the model.

[@ayeshasiddiqui9068]

Please make a video on phd viva questions